LoraCTL is highly suggested and CFG above 7 is fine.
Description
FAQ
Comments (8)
Hi fruitmax, been using all of your loras so far, I would like to ask, is the trigger word necessary in your loras? If so, I find that the way you set your trigger word is very inconvenient for the users, I understand the meaning but it is just very unintuitive. This one is also on the more extreme side where the lora filename is almost unrecognizable unless we rename it ourselves when we download it. Would you consider just use the normal pose/person name for the tigger word from now on?
The problem with that is that you get better results with a unique label, hence the words with letters replaced by numbers. You also don't necessarily need to place the trigger word if you describe the image correctly.
However adding the trigger word will force the image to fit the image "style" of the reference images. Otherwise the image style will resemble more of the starting checkpoint at the risk of body morphing and other misbehaviors.
https://civitai.com/articles/3105/essential-to-advanced-guide-to-training-a-lora
When introducing a new concept, Dreambooth suggests using uncommon or rare tokens - unique identifiers the model doesn’t strongly associate with anything it already knows. Because of their weak prior associations, it will make the language drift less impactful (since it will lose an association that probably made little sense). This prevents the new training from interfering with the model’s existing capabilities. Examples of such tokens might be: "olis", "bnha", or "hta". To represent your new concepts with a rare token, you might then use the rare token with the class (e.g. "a olis girl", where olis could be any new character you would like to train).
I see, thanks for clearing that up, keep up the good work.
Wow, this is the most responsive and well made lora I've seen for this position. Absolutely FANTASTIC WORK!
People tend to stay with Loras even if I find Lycoris to be far superior/flexible as you said. They also requires less space for the same LORA complexity. With my current hardware, a 77.8 MB Lycoris model takes about an hour, if I wanted to make an equivalent LORA model I would need to pump up the complexity of the model to double its size and would require double the training time...
If it's not too much trouble, would you consider making a POV version of this (were the woman is riding the viewer directly)???
It's to be used with loractl extension in automatic1111. Basically this activate fully the lora at the beginning of the diffusion process, then at 80% it starts to drop. This way you can let the checkpoint do the finer details.